400 error: `reasoning_content` must be passed back for DeepSeek-family models on direct OpenAI-compatible endpoints

[Timmy6942025]

Problem

Multi-turn sessions against DeepSeek-family models served by direct OpenAI-compatible profiles (no OpenRouter provider features — e.g. OpenCode Zen, opencode.ai/zen/v1) fail on the second request with:

400 The `reasoning_content` in the thinking mode must be passed back to the API.

The endpoint runs server-managed thinking mode and requires the assistant reasoning_content to be replayed on follow-up requests. jcode stores it (stores_reasoning_content_for_context) but drops it at request-build time for these profiles, because reasoning passback is only unlocked when OpenRouter provider features are enabled or the endpoint is the Kimi coding endpoint (#322).

Repro

• Any model in the DeepSeek family (e.g. deepseek-v4-flash-free on OpenCode Zen) via a direct OpenAI-compatible profile
• Start a session, let the model emit a tool call (which carries a stored reasoning block), then send a follow-up request
• 400 error above

Workaround (verified): set allow_provider_pinning = true on the named profile so provider features are treated as enabled, which unlocks the reasoning passback path.

Related

• Kimi API fails with tool calls when thinking mode is enabled #322 — same mechanism for the Kimi coding endpoint (fixed)
• After jcode is actually doing something, gets interrupted, and then executes another task, it returns a 400 error #321, Deepseek models: deepseek-ai/deepseek-v4-flash, deepseek-ai/deepseek-v4-pro for Nvidia Nim provider not working #341 — neighboring DeepSeek/Kimi thinking-mode 400s

[woodssusan98002]

👍 索引设计踩过什么坑

[harringtonjoshua7584]

讨论区比 Issue 更适合提问吗 👍

[Timmy6942025]

Ready-to-apply patch (verified against master 02439b4, fully tested). The change is on branch fix/deepseek-reasoning-content-passback in fork Timmy6942025/jcode (commit 4e9f6c9) — a PR couldn't be opened from this account due to a GitHub-side permission flag on PR creation, so here's everything in one place.

Production change

crates/jcode-provider-openrouter-runtime/src/openrouter_provider_impl.rs, in complete():

@@ -53,6 +53,18 @@
         // `thinking` field (the endpoint already manages thinking itself), so
         // this is kept separate from `thinking_enabled`.
         let kimi_coding_endpoint = self.is_kimi_coding_endpoint(&model);
+        // DeepSeek-family models served through direct OpenAI-compatible
+        // endpoints (e.g. OpenCode Zen) run in server-managed thinking mode and
+        // require `reasoning_content` to be passed back on assistant messages in
+        // multi-turn requests; DeepSeek's API docs mandate it and the endpoint
+        // enforces it with a 400 "The `reasoning_content` in the thinking mode
+        // must be passed back to the API." error. Those profiles carry no
+        // OpenRouter provider features, so unlock reasoning passback by model
+        // family instead. Like the Kimi coding endpoint, this must NOT add the
+        // OpenRouter-specific top-level `thinking` field (the endpoint manages
+        // thinking itself), so it stays separate from `thinking_enabled`.
+        let deepseek_thinking_endpoint =
+            !self.supports_provider_features && Self::model_is_deepseek_family(&model);
         let thinking_enabled = thinking_override.or_else(|| {
             if Self::is_kimi_model(&model) {
                 Some(true)
@@ -60,8 +72,9 @@
                 None
             }
         });
-        let allow_reasoning = (self.supports_provider_features || kimi_coding_endpoint)
-            && thinking_enabled != Some(false);
+        let allow_reasoning =
+            (self.supports_provider_features || kimi_coding_endpoint || deepseek_thinking_endpoint)
+                && thinking_enabled != Some(false);
         let include_reasoning_content = thinking_enabled == Some(true)
             || (allow_reasoning && Self::is_kimi_model(&model))
             || kimi_coding_endpoint;

This mirrors the existing Kimi coding-endpoint handling (#322) and deliberately does not add the OpenRouter-specific top-level thinking field — the endpoint manages thinking itself.

Regression test

crates/jcode-provider-openrouter-runtime/src/openrouter_deepseek_tests.rs (new file, registered as mod openrouter_deepseek_tests in lib.rs): deepseek_direct_compat_tool_call_message_replays_stored_reasoning_content — builds an OpenCode-Zen-style direct profile (supports_provider_features: false), replays a multi-turn conversation whose assistant tool-call turn carries a stored reasoning block, and asserts the outgoing request replays reasoning_content on the assistant message.

Verification

• cargo test -p jcode-provider-openrouter-runtime: 113 passed, 0 failed (regression test included)
• cargo fmt --all -- --check: clean
• Clippy: no new diagnostics from changed files
• Test/other ratchets: no new violations vs master
• Live-verified end-to-end: multi-turn tool loop against OpenCode Zen works with the fix (also reproducible via config workaround allow_provider_pinning = true on the named profile)

Files

• crates/jcode-provider-openrouter-runtime/src/openrouter_provider_impl.rs (+13)
• crates/jcode-provider-openrouter-runtime/src/openrouter_deepseek_tests.rs (+112, new)
• crates/jcode-provider-openrouter-runtime/src/lib.rs (+4, test-module registration)
• crates/jcode-provider-openrouter-runtime/src/openrouter_tests.rs (±0 net: 3 helpers made pub(crate))

Reference PR body (in case PR creation becomes possible): master...Timmy6942025:fix/deepseek-reasoning-content-passback